📦 zhayujie/CowAgent：开源超级 AI 助手与 Agent 运行框架

[English] | [中文] | [日本語]

CowAgent 是一个开源的超级 AI 助手，它能主动规划任务、控制你的计算机和外部服务、创建并运行技能，构建个人知识库和长期记忆，并通过自我进化与你共同成长——这也是 Agent Harness 工程的一个参考实现。

CowAgent 轻量、易于部署，并且高度可扩展。你可以接入任意主流 LLM 供应商，并在个人电脑或服务器上 7×24 小时运行它，覆盖 Web 和所有主流即时通讯平台。

🌐 官网 · 📖 文档 · 🚀 快速开始 · 🧩 技能中心 · ☁️ 在线体验

能力	描述
规划	将复杂任务分解并逐步执行，循环调用工具直至达成目标
记忆	三层架构（上下文 → 日常 → 核心），自动 Deep Dream 蒸馏，混合关键词 + 向量检索
知识	自动将结构化知识整理成 Markdown 维基，构建不断演进的知识图谱，并支持可视化浏览
进化	自我进化自动回顾对话，以改进技能、跟进未完成的任务，并巩固记忆和知识，在日常使用中不断成长
技能	从技能中心、GitHub、ClawHub 一键安装；或通过自然语言对话创建自定义技能
工具	内置文件 I/O、终端、浏览器、调度器、记忆检索、网络搜索等 10 多种工具——并原生集成 MCP 协议
频道	集成 Web、微信、飞书、钉钉、企业微信、QQ、公众号、Telegram 和 Slack
多模态	一流支持文本、图片、语音和文件——识别、生成与投递
模型	Claude、GPT、Gemini、DeepSeek、Qwen、GLM、Kimi、MiniMax、豆包等——可在 Web 控制台一键切换供应商
部署	一行命令安装器，统一的 Web 控制台，多种部署模式（本地、Docker、服务器）

CowAgent 是一个完整的 Agent Harness：消息通过频道流入；智能体核心根据记忆、知识、可用工具和技能进行规划与推理；模型生成响应，并通过原始频道返回。每一层都是解耦且独立可扩展的。

了解更多请参阅架构文档。

一行命令安装器会处理依赖、配置和启动：

Linux / macOS:

bash <(curl -fsSL https://cdn.link-ai.tech/code/cow/run.sh)

Windows (PowerShell):

irm https://cdn.link-ai.tech/code/cow/run.ps1 | iex

Docker:

curl -O https://cdn.link-ai.tech/code/cow/docker-compose.yml
docker compose up -d

启动后，在浏览器中打开 http://localhost:9899 即可访问 Web 控制台——这是一个一站式中心，你可以在这里与 Agent 对话、配置模型、连接频道和安装技能。

如果你部署在服务器上？在 config.json 中将 web_host 设置为 0.0.0.0 以便从外部访问控制台，并设置 web_password 进行保护。不要忘记在防火墙或安全组中开放端口 9899。

📖 详细指南：快速开始 · 源码安装 · 升级

安装后，你可以使用 cow CLI 管理服务：

cow start | stop | restart        # 服务控制
cow status | logs                  # 状态与日志
cow update                         # 拉取最新代码并重启
cow skill install <name>           # 安装技能
cow install-browser                # 安装浏览器自动化

CowAgent 支持所有主流 LLM 供应商。聊天、视觉、图像生成、ASR/TTS 和嵌入 可以分别路由到不同的供应商。供应商配置可直接在 Web 控制台完成，无需手动编辑文件。

供应商	重点模型	聊天	视觉	图像生成	ASR	TTS	嵌入
Claude	claude-sonnet-5	✅	✅
OpenAI	gpt-5.5, o-series	✅	✅	✅	✅	✅	✅
Gemini	gemini-3.5-flash	✅	✅	✅
DeepSeek	deepseek-v4-flash / pro	✅
Qwen	qwen3.7-plus	✅	✅	✅	✅	✅	✅
GLM	glm-5.2, glm-5v-turbo	✅	✅		✅		✅
豆包	doubao-seed-2.1 系列	✅	✅	✅			✅
Kimi	kimi-k2.7-code	✅	✅
MiniMax	MiniMax-M3	✅	✅	✅		✅
ERNIE	ernie-5.1	✅	✅
MiMo	mimo-v2.5 / pro	✅	✅			✅
LinkAI	一键接入 100+ 模型	✅	✅	✅	✅	✅	✅
自定义	本地模型 / 第三方代理	✅

关于每个供应商的详细信息，请参阅模型概览。

单个 Agent 实例可以并行服务多个频道。大多数频道都可以直接从 Web 控制台完成接入。

请参阅频道概览了解设置详情。

Web 控制台是默认频道，也是配置模型、频道、技能、记忆等的统一入口。

🧠 记忆与知识库

长期记忆采用三层架构：会话上下文（短期）→ 日常记忆（中期）→ MEMORY.md（长期）。每晚一次的 Deep Dream 过程会将分散的记忆提炼为精炼的长期条目和叙事日志。请参阅长期记忆 · Deep Dream。

个人知识库通过按主题组织结构化知识，对时间顺序的记忆进行补充。Agent 会自动从对话中收集有价值的信息，维护交叉引用和索引，Web 控制台还提供了交互式知识图谱视图。请参阅个人知识库。

长期记忆 · 三层架构 + Deep Dream	知识库 · 自动整理的 Markdown 维基

工具是 Agent 用来与系统资源交互的原子能力。技能则是由清单文件定义的更高层次的工作流，可将多个工具组合起来完成复杂任务。

内置工具涵盖文件 I/O (read / write / edit / ls)、终端 (bash)、文件发送 (send)、记忆检索 (memory)、环境变量 (env_config)、网页抓取 (web_fetch)、调度 (scheduler)、网络搜索 (web_search)、视觉 (vision) 和浏览器自动化 (browser)。

MCP 协议集成了开放的 Model Context Protocol 服务器生态。只需一个 mcp.json 文件——支持 stdio / SSE 传输、热重载和零代码集成。

了解更多：工具概览 · MCP 集成。

• 技能中心——开放技能市场：浏览、搜索、一键安装
• GitHub / ClawHub / URL 等多种来源——从任意来源安装技能

[English] | [中文] | [日本語]

CowAgent is an open-source super AI assistant that proactively plans tasks, controls your computer and external services, creates and runs Skills, builds a personal knowledge base and long-term memory, and grows alongside you through self-evolution — a reference implementation of Agent Harness engineering.

CowAgent is lightweight, easy to deploy, and built to extend. Plug in any major LLM provider and run it 24/7 on a personal computer or server, across the web and all major IM platforms.

🌐 Website · 📖 Docs · 🚀 Quick Start · 🧩 Skill Hub · ☁️ Try Online

Capability	Description
Planning	Decomposes complex tasks and executes them step by step, looping over tools until the goal is reached
Memory	Three-tier architecture (context → daily → core), automatic Deep Dream distillation, hybrid keyword + vector retrieval
Knowledge	Auto-curates structured knowledge into a Markdown wiki, builds an evolving knowledge graph with visual browsing
Evolution	Self-Evolution reviews conversations automatically to improve skills, follow up on unfinished tasks, and consolidate memory and knowledge, growing through everyday use
Skills	One-click install from Skill Hub, GitHub, ClawHub; or create custom skills via natural-language conversation
Tools	Built-in file I/O, terminal, browser, scheduler, memory retrieval, web search, and 10+ more tools — with native MCP integration
Channels	Integrates with Web, WeChat, Feishu, DingTalk, WeCom, QQ, Official Accounts, Telegram, and Slack
Multimodal	First-class support for text, images, voice, and files — recognition, generation, and delivery
Models	Claude, GPT, Gemini, DeepSeek, Qwen, GLM, Kimi, MiniMax, Doubao, and more — swap providers from the Web console with one click
Deploy	One-line installer, unified Web console, multiple deployment modes (local, Docker, server)

CowAgent is a complete Agent Harness : messages flow in through Channels ; the Agent Core plans and reasons over memory, knowledge, and the available tools and skills; Models generate the response, which is sent back through the originating channel. Every layer is decoupled and independently extensible.

Read more in Architecture.

A one-line installer takes care of dependencies, configuration, and startup:

Linux / macOS:

bash <(curl -fsSL https://cdn.link-ai.tech/code/cow/run.sh)

Windows (PowerShell):

irm https://cdn.link-ai.tech/code/cow/run.ps1 | iex

Docker:

curl -O https://cdn.link-ai.tech/code/cow/docker-compose.yml
docker compose up -d

Once started, open http://localhost:9899 to access the Web console — your one-stop hub to chat with the Agent, configure models, connect channels, and install skills.

Deploying on a server? Set web_host to 0.0.0.0 in config.json to make the console reachable from outside, and set web_password to protect it. Don't forget to open port 9899 in your firewall or security group.

📖 Detailed guides: Quick Start · Install from Source · Upgrade

After installation, manage the service with the cow CLI:

cow start | stop | restart        # service control
cow status | logs                  # status and logs
cow update                         # pull latest code and restart
cow skill install <name>           # install a skill
cow install-browser                # install browser automation

CowAgent supports all mainstream LLM providers. Chat, vision, image generation, ASR/TTS, and embeddings can each be routed to a different vendor. Providers are configured directly in the Web console — no manual file editing required.

Provider	Featured Models	Chat	Vision	Image Gen	ASR	TTS	Embedding
Claude	claude-sonnet-5	✅	✅
OpenAI	gpt-5.5, o-series	✅	✅	✅	✅	✅	✅
Gemini	gemini-3.5-flash	✅	✅	✅
DeepSeek	deepseek-v4-flash / pro	✅
Qwen	qwen3.7-plus	✅	✅	✅	✅	✅	✅
GLM	glm-5.2, glm-5v-turbo	✅	✅		✅		✅
Doubao	doubao-seed-2.1 series	✅	✅	✅			✅
Kimi	kimi-k2.7-code	✅	✅
MiniMax	MiniMax-M3	✅	✅	✅		✅
ERNIE	ernie-5.1	✅	✅
MiMo	mimo-v2.5 / pro	✅	✅			✅
LinkAI	One key for 100+ models	✅	✅	✅	✅	✅	✅
Custom	Local models / third-party proxy	✅

For details on each provider, see the Models overview.

A single Agent instance can serve multiple channels in parallel. Most channels can be onboarded right from the Web console.

See the Channels overview for setup details.

The Web console is the default channel and the unified entry point to configure models, channels, skills, memory, and more.

🧠 Memory & Knowledge Base

Long-term memory uses a three-tier architecture: conversation context (short-term) → daily memory (mid-term) → MEMORY.md (long-term). A nightly Deep Dream pass distills scattered memories into refined long-term entries and a narrative journal. See Long-term Memory · Deep Dream.

Personal knowledge base complements the time-ordered memory by organizing structured knowledge by topic. The Agent automatically curates valuable information from conversations, maintains cross-references and indexes, and the Web console offers an interactive knowledge-graph view. See Personal Knowledge Base.

Long-term Memory · Three-tier architecture + Deep Dream	Knowledge Base · Auto-curated Markdown wiki

Tools are atomic capabilities the Agent uses to interact with system resources. Skills are higher-level workflows defined by a manifest file that compose multiple tools to accomplish complex tasks.

Built-in tools cover file I/O (read / write / edit / ls), terminal (bash), file sending (send), memory retrieval (memory), environment variables (env_config), web fetching (web_fetch), scheduling (scheduler), web search (web_search), vision (vision), and browser automation (browser).

MCP protocol integrates the open ecosystem of Model Context Protocol servers. A single mcp.json is enough — supports stdio / SSE transports, hot reload, and zero-code integration.

Learn more: Tools overview · MCP integration.

• Skill Hub — open skill marketplace: browse, search, install in one click

• GitHub / ClawHub / URL and more — install skills from any source

• Conversational authoring — generate custom skills through dialogue with skill-creator; turn any workflow or third-party API into a reusable skill

  /skill list # list installed skills /skill search # search the marketplace /skill install # one-click install

Learn more: Skills overview · Creating Skills.

2026.06.18: v2.1.2 — Web console upgrades (scheduled task management, knowledge base categories, multiple custom model providers), Self-Evolution improvements, new models (kimi-k2.7-code, glm-5.2), security hardening and refinements.

2026.06.09: v2.1.1 — Self-Evolution, Web console upgrades (message management, parallel sessions), cross-platform MCP enhancements with concurrent calls, new models (MiniMax-M3, qwen3.7-plus), Python 3.13 support.

2026.06.01: v2.1.0 — Internationalization, new channels (Telegram, Discord, Slack, WeChat Customer Service), CLI interaction upgrades, streamlined one-line install, MCP Streamable HTTP support, new models (claude-opus-4-8, MiMo).

2026.05.22: v2.0.9 — Model management, MCP protocol support, persistent browser sessions, new models (gpt-5.5, gemini-3.5-flash, qwen3.7-max), deployment hardening.

2026.05.06: v2.0.8 — Feishu channel overhaul (voice, streaming, QR onboarding), DeepSeek V4 and Baidu Qianfan support, scheduler tool upgrades.

2026.04.22: v2.0.7 — Built-in image generation (GPT Image 2, Nano Banana), new models (Kimi K2.6, Claude Opus 4.7, GLM 5.1), memory and knowledge enhancements.

2026.04.14: v2.0.6 — Knowledge base, Deep Dream memory distillation, smart context compression, multi-session Web console.

2026.04.01: v2.0.5 — Cow CLI, Skill Hub open source, browser tool, WeCom Bot QR onboarding.

2026.02.03: v2.0.0 — Major upgrade to a super Agent assistant with multi-step task planning, long-term memory, and the Skills framework.

Full history: Release Notes

🤝 Community & Support

File an issue on GitHub, or scan the QR code below to join our WeChat community:

• Cow Skill Hub — open skill marketplace for AI Agents; works with CowAgent, OpenClaw, Claude Code, and more
• bot-on-anything — lightweight LLM application framework with integrations for Slack, Telegram, Discord, Gmail, and more
• AgentMesh — open-source multi-agent framework for solving complex problems through team collaboration

LinkAI is an all-in-one AI Agent platform for enterprises and developers, offering managed hosting and enterprise-grade support for CowAgent:

• 🚀 Zero-deployment hosted runtime — spin up a CowAgent online assistant in under a minute, no server required
• 🧠 Agent infrastructure — unified access to LLMs, knowledge bases, databases, skills, and workflows; plug-and-play building blocks that extend what CowAgent can do
• 🏢 Team & enterprise features — workspaces, role-based access, audit logs, and private deployment for production use cases

For enterprise inquiries: sales@simple-future.tech or scan the QR code to reach our team on WeChat.

🛠️ Development & Contributing

All kinds of contributions are welcome — new features, bug fixes, performance improvements, docs, or sharing your own skills on the Skill Hub. See CONTRIBUTING.md to get started, then open an Issue to discuss or send a PR directly.

⭐ Star the project to show your support, and Watch → Custom → Releases to get notified of new versions. PRs and Issues are always welcome.

1. This project is licensed under the MIT License and is intended for technical research and learning. You are responsible for complying with applicable laws and regulations in your jurisdiction; the maintainers assume no liability for any consequences arising from use of this project.
2. Cost & safety: Agent mode consumes substantially more tokens than regular chat — pick models that balance quality and cost. The Agent has access to your local operating system, so only deploy it in trusted environments.
3. CowAgent is a pure open-source project and does not participate in, authorize, or issue any cryptocurrency.

📌 Project Renaming Notice

This project was previously named chatgpt-on-wechat and is now officially CowAgent. The old GitHub URL redirects automatically; existing users may optionally run git remote set-url origin https://github.com/zhayujie/CowAgent.git to update the local remote.